Joint sparse representation and deep learning-based image super resolution method and system

Abstract

The invention relates to a joint sparse representation and deep learning-based image super resolution method and a joint sparse representation and deep learning-based image super resolution system. The method includes the following steps that: resolution reduction is performed on an original high-resolution image, so that a low-resolution image of which the size is the same as the original high-resolution image, and the difference value part of the original high-resolution image and the low-resolution image is obtained; a low-resolution image dictionary, a difference value image dictionary and corresponding sparse representation coefficients are obtained; a deep learning network with root-mean-square error adopted as a cost function is constructed, and network parameters are optimized iteratively, so that the cost function can be minimum, a trained deep learning network can be obtained; and the sparse coefficient of the low-resolution image, which is adopted as a test part, is inputted into the deep learning network, when error is smaller than a given threshold value, a corresponding high-resolution image can be reconstructed according to the low-resolution image of which the resolution is to be improved. According to the method and system of the invention, defects of an existing method according to which a joint dictionary training mode is utilized to make a high-resolution image and a low-resolution image share a sparse representation coefficient can be eliminated, and deep learning is utilized to fully learn the mapping relationship between the low-resolution image sparse representation coefficient and the difference value image sparse representation coefficient, and therefore, a high-resolution reconstruction result with higher precision can be obtained.

Description

technical field [0001] The invention belongs to the technical field of remote sensing image processing, and relates to an image super-resolution method and system for joint sparse representation and deep learning. Background technique [0002] Image super-resolution is to reconstruct a high-resolution image containing more details from a series of lower-resolution images. It has important application value in target recognition and positioning of remote sensing images, environmental monitoring, medical imaging and many other fields. . Image super-resolution breaks through the limitation of the resolution of the sensor itself, and obtains images with higher quality and higher resolution based on the existing image acquisition technology, which provides a basis for further image analysis. [0003] Traditional image super-resolution requires multiple low-resolution images of the same scene as information sources. Based on reasonable assumptions and prior information for mappin...

AI Technical Summary

Problems solved by technology

Freeman et al. used Markov random field to establish the correspondence between high and low image blocks, and used the Bayesian confidence algorithm to solve it; Change et al. applied the idea of ​​domain embedding in manifold learning to image super-resolution, assuming that LR image blocks It is similar to the local manifold of the HR image block, by solving the k-neighborhood representation coefficients of the test image block in the manifold formed by the low-resolution image block, and finally using these coefficients to linearize the k-neighborhood of the HR image block Combined to get the corresponding HR image blocks, but this method is also prone to underfitting or overfitting; Yang et al. introduced some ideas of compressed sensing into the super-resolution algorithm, and through linear combination with HR image blocks get high resolution images

Images